Main product detection with graph networks for fashion

Abstract Computer vision has established a foothold in the online fashion retail industry. Main product detection is crucial step of vision-based feed parsing pipelines, focused on identifying bounding boxes that contain being sold gallery images page. The current state-of-the-art approach does not leverage relations between regions image, and treats same independently, therefore fully exploiting visual contextual information. In this paper, we propose model incorporates Graph Convolutional Networks (GCN) jointly represent all detected as nodes. We show proposed method better than state-of-the-art, especially, when consider scenario where title-input missing at inference time for cross-dataset evaluation, our outperforms previous approaches by large margin.